V3-V4 info: Two more versions trained with more recent features. Both was trained with T5 and clip-l encoders. V3 was trained with long prose from joycaption. V4 was trained with trimmed captions from torii captioner. Between those two I personally prefer V3.
V1-V2 info: Two versions, first one is trained with 512 res and booru captions for 2000 steps and second is 768 res trained for 5000 steps with nlp captions. Can't say which one of them is strictly better, probably second one, since it's transferring style more harder even with another char lora. Haven't tested it with anything else than base checkpoint in fp16/fp8 mode and afaik rn it will not work with some quantization types in popular ui's.
Workflows with prompts are right inside the images, you can just drag them into your comfyui browser window, probably could work better with "anime screencap" in pos and "realistic, 3d" in negs", negs unavail with 1 cfg, but you could hack it with this extension, settings for which is also into the flow already.
Description
Version trained with encoders and full llm slop prose captions from joy caption at 768 res.
FAQ
Comments (6)
Wow, you have so perfect and fine-tuned loras TwwT Can you tell how you train your Loras? How much images you need to train styles for Flux? How steps?
You did nice job!
Thanks. Same as for any other arch, 100-200 should be fine, steps varied between versions, for the last one it was 3000, I can drop easy-scripts config if you want it for any specific version
@bakariso If you okay with it, it will be super cool to get script mb for 3 version. IF you don't mind of course!
@DualChimerra For the latest two I haven't used any kind of gui with configs, raw command line is everything I can share here, obviously you should change everything "CHANGE" to your own fs paths. It's contradictional about training encoder or not, I was just testing how it would turn out with that and recommended params from sd-scripts repo. I'm pretty sure if you read sd-scripts discussion about flux there is better algo such as lokr, better optimizer such as ademamix and even better based undistilled checkpoint to start with, but I haven't tested almost all of that, only ademamix and didn't like the result at the end. Here is the cmd, use it after getting into venv https://files.catbox.moe/u53hs8.txt hope that helps somehow
@DualChimerra Oh and there is alot of fuckups in that cmd that I haven't even noticed, like min_snr_gamma shouldn't be there, when there is already timestep shift being applied. Or clip_skip, at least doing nothing in that case
@bakariso Thanks you for sharing with me ! Have a nice day :)